Hi Urs,
thanks for the interesting summary.
The "problem" cannot be caused by annotated junctions as they are present in the reference for both 1st and 2nd path.
The problem is most likely caused by those novel junctions that are detected at the 1st pass with low to moderate expression, but overlap the highly expressed loci. In this case, the many reads that map to these loci unspliced in the 1st pass, in the 2nd pass have to be "checked" for splicing, and I think this is what causes slowdown and increase in multimappers.
In your example, in Sample2, the chr14 (IG locus?) stands out compared to"good" Samples1/3, in large numbers of unannotated junctions. The fact that these junctions get high counts indicates that these loci are highly expressed.
Similarly, Sample4 stands out in chr14 and chr2.
Now, the question you need to answer is whether 2-pass alignment really adds any value for your downstream analyses. Note that 2-pass main effect is to increase the number of reads mapping to unannotated junctions. If you are not interested that much in unannotated junctions, you may simply use 1-pass runs with annotations.
If you still want to get more accurate counts for unannotated junctions, you would need to filter out 1st pass novel junctions that map to highly expressed loci on chr14 and chr2. After that, you can run the 2nd pass "manually" with --sjdbFileChrStartEnd filtered.SJ.out.tab .
Cheers
Alex